AI BriefingHugging Face Model Security Scanning Achievements
Key point
Hugging Face and Protect AI collaborated to scan 4.47 million models and detect over 350,000 security threats.
Details
The partnership between Hugging Face and Protect AI marks its 6-month anniversary with an announcement of results. Protect AI's Guardian technology is being used to strengthen model security on the Hugging Face Hub.
New Detection Modules Released:
- PAIT-ARV-100: Detects file system write threats when loading compressed files
- PAIT-JOBLIB-101: Detects suspicious code execution when loading Joblib models
- PAIT-TF-200: Detects backdoors in TensorFlow SavedModel architectures
- PAIT-LMAFL-300: Detects malicious code execution during Llamafile inference
- Includes detection of a high-risk vulnerability in Keras (CVE-2025-1550)
Key Achievements (as of April 1, 2025):
- Completed scanning of 4.47 million unique model versions and 1.41 million repositories
- Identified a total of 352,000 security issues across 51,700 models
Protect AI applies a Zero Trust approach, treating arbitrary code execution as a potential threat, and immediately reflects over 200 vulnerability reports collected through the huntr bug bounty program into Guardian.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Tailscale Didn't Stop the Hugging Face Intrusion